测试视频复刻

core/backend/apps/ai/video_digest.py

@@ -48,6 +48,8 @@ MAX_FRAMES = 36
 FRAME_WIDTH = 768
 FRAME_QUALITY = 3
 DIGEST_MAX_TOKENS = 12288
+# 视频复刻内联拆解的同模型重试次数(Gemini 偶尔吐废稿;复刻失败要用户从头重选素材,值得多试一次)
+DIGEST_MAX_ATTEMPTS = 2
 _SHOT_MARK = re.compile(r"【(?:镜头\s*\d+|第\s*\d+\s*镜)】")
 
 _FFMPEG_TIMEOUT = 60
@@ -728,12 +730,12 @@ def _store_raw_source(*, team, path: str, suffix: str) -> tuple[str, str]:
     return stored.object_key, storage.public_url(object_key=stored.object_key)
 
 
-def _download_source_to_temp(object_key: str, suffix: str) -> str:
+def _download_source_to_temp(object_key: str, suffix: str, *, bucket: str = "") -> str:
     from apps.assets.storage import TosStorage
 
     ext = suffix if str(suffix).startswith(".") else f".{suffix or 'mp4'}"
     storage = TosStorage()
-    body = storage.client.get_object(Bucket=storage.bucket, Key=object_key)["Body"].read()
+    body = storage.client.get_object(Bucket=bucket or storage.bucket, Key=object_key)["Body"].read()
     tmp = tempfile.NamedTemporaryFile(suffix=ext.lower(), delete=False)
     try:
         tmp.write(body)
@@ -1357,7 +1359,7 @@ def digest_asset_video(*, asset, task, model_config_id=None) -> tuple[str, dict]
     复刻本来就是一次收费,拆解是它的内部工序。模型调用的审计仍记在传入的复刻
     task 上(AIModelAttempt),出问题查得到。
     """
-    from apps.ai.services import execute_routed_text_request
+    from apps.ai.services import _collect_extract_text, get_text_provider
 
     primary = asset.files.filter(is_primary=True).first() or asset.files.first()
     if primary is None or not primary.object_key:
@@ -1372,7 +1374,7 @@ def digest_asset_video(*, asset, task, model_config_id=None) -> tuple[str, dict]
 
     local_path = ""
     try:
-        local_path = _download_source_to_temp(primary.object_key, suffix)
+        local_path = _download_source_to_temp(primary.object_key, suffix, bucket=primary.bucket or "")
         base_name = (asset.name or "参考视频").rsplit(".", 1)[0]
         upload = _FileFromPath(local_path, f"{base_name}{suffix}")
         video, frames, duration, extras = digest_input_from_upload(upload)
@@ -1385,6 +1387,11 @@ def digest_asset_video(*, asset, task, model_config_id=None) -> tuple[str, dict]
             duration,
             len(frames),
         )
+        # ⚠️ 这里必须和「提炼提示词」页(_digest_video)构造出完全相同的 messages:
+        # 同一份 SKILL.md 作 system、同一段 user 引导语、同样的 temperature / max_tokens。
+        # 绝对不要在这里追加商品信息(product_hint 保持默认空)——两处提炼稿一旦不同,
+        # 用户在提炼页看到的分镜和复刻实际用的分镜就对不上,排查会乱。
+        # test_video_replace.test_replace_digest_prompt_matches_standalone 会守住这一点。
         messages = build_digest_messages(
             frames,
             duration,
@@ -1392,28 +1399,34 @@ def digest_asset_video(*, asset, task, model_config_id=None) -> tuple[str, dict]
             aspect_ratio=ratio_label(int(extras.get("width") or 0), int(extras.get("height") or 0)),
             file_title=title_from_filename(str(extras.get("file_name") or "")),
         )
-        routed = execute_routed_text_request(
-            task=task,
-            primary_model=model_config,
-            messages=messages,
-            streaming=True,
-            structured_output=False,
-            business_operation="video_digest",
-            temperature=0.4,
-            validate_text=lambda text: validate_digest_text(
-                text, duration=duration, frame_count=len(frames) or (24 if video else 0)
-            ),
-            extra_body={"max_tokens": DIGEST_MAX_TOKENS},
-            request_summary={
-                "duration_seconds": round(duration, 2),
-                "frame_count": len(frames),
-                "input": "native_video" if video is not None else "frames",
-                "for": "video_replace",
-            },
-            allow_retry=False,
-            allow_fallback=False,
-        )
-        _text, _response, digest = routed.value
+        # 这里不能走 execute_routed_text_request:它硬性要求 task.model_config 就是本次主模型,
+        # 而复刻任务的 model_config 是 Seedance(视频模型),传进去必抛
+        # 「AITask.model_config 必须保持为用户选择或系统默认的主模型」。
+        # 改为直连脚本/提炼同一条流式通道,并自己做「同模型重试」——绝不 fallback 到别的
+        # 文本模型,它们看不了视频,换过去必然废稿。
+        provider = get_text_provider(model_config)
+        expected_frames = len(frames) or (24 if video else 0)
+        digest = ""
+        last_error: Exception | None = None
+        for attempt in range(1, DIGEST_MAX_ATTEMPTS + 1):
+            try:
+                text, _payload = _collect_extract_text(
+                    provider,
+                    model_config,
+                    messages,
+                    temperature=0.4,
+                    extra_body={"max_tokens": DIGEST_MAX_TOKENS},
+                )
+                digest = validate_digest_text(text, duration=duration, frame_count=expected_frames)
+                break
+            except Exception as exc:  # noqa: BLE001 — 空文/废稿/网络抖动都值得再来一次
+                last_error = exc
+                logger.warning(
+                    "replace digest attempt %s/%s failed for task %s: %s",
+                    attempt, DIGEST_MAX_ATTEMPTS, task.id, exc,
+                )
+        if not digest:
+            raise VideoDigestError(f"参考视频拆解失败：{last_error}")
         meta = {
             "digest_model": model_config.name,
             "digest_input": "native_video" if video is not None else "frames",